siegfly/dsh-deepseek-vision
A vision-language gateway provider route: pasted images are described by a configurable VL model (Qwen-VL by default) before the DeepSeek wire.
dsh-deepseek-vision is a vision-language gateway plugin: models the DeepSeek catalog declares image-capable (the official deepseek-v4-flash-vision-exp) stream images natively to DeepSeek's vision endpoint, while every other (text-only) model gets pasted images described verbatim by a configurable VL model (Qwen-VL by default) before reaching the text-only wire. It positions itself as the thinnest bridge among dsh vision plugins — no injected agent tools, no third-party relay, no local model requirement — and locks nothing across versions. Descriptions are cached with an in-process LRU keyed on the full VL configuration generation plus attachmentId, so retries, compaction and later turns reuse the same description without double billing; original images stay persisted in the session log so history/replay are unaffected. Fail-closed by default with stable error codes, or a placeholder degrade mode; any OpenAI-style /chat/completions gateway (DashScope, vLLM, OpenRouter, LM Studio) works via a settings card.
Install
dsh plugin --profile web add dsh-deepseek-visionThe README's install line is 'dsh plugin --profile web add dsh-deepseek-vision'; if dsh is not on PATH it documents the npx form 'npx @deepseek-ai/dsh plugin --profile web add dsh-deepseek-vision'. The npm package dsh-deepseek-vision 0.1.7 was re-verified live on 2026-10-02 with its repository field back-linking to siegfly/dsh-deepseek-vision (MIT; release v0.1.7, 121 tests, 97% coverage). Four spec forms (npm / git / directory / tarball) and web + headless profiles are documented.
Compatibility
DeepSeek Harness web and headless. Node.js ^22.19 || >=24 (only for the no-CLI replica path). Registers the deepseek-vision route (display name 'DeepSeek + Vision') with a real inputModalities ['text','image'] declaration, so chat pastes, tool-fs read_image, browser screenshots, MCP tool images and ACP inline images get through.
Details
- Repo: siegfly/dsh-deepseek-vision
- Category: Vision & Multimodal
- Stars: 0
- Version: npm dsh-deepseek-vision 0.1.7 (registry re-verified 2026-10-02); MIT
- Last push: 2026-09-11
- First seen: 2026-08-15
Recent updates
The README documents the highlights, the 'Why This One' comparison table against tool-style and router-style vision plugins, the configuration section, four install spec forms, the version-alignment section (no cross-version lock-in), and the FAQ.
FAQ
- How do I install dsh-deepseek-vision?
- dsh plugin --profile web add dsh-deepseek-vision (npm 0.1.7 re-verified 2026-10-02); use the npx @deepseek-ai/dsh prefix if the CLI is not on PATH.
- Does it work with a text-only DeepSeek model?
- Yes — models the catalog declares image-capable stream images natively to the DeepSeek vision endpoint; other models get a verbatim VL description first, with no model swap.
- Which VL providers can it use?
- Any OpenAI-style /chat/completions gateway — the README names DashScope, vLLM, OpenRouter and LM Studio, with the VL model swap done in the settings card.
Alternatives
gloryxpnv/dsh-tool-vision · welsione/dsh-mmx-bridge · jiavenzhong/dsh-tool-markitdown